독서

Intel's 18A process node has achieved logic density comparable to TSMC's N3E, but does not yet lead TSMC's more advanced N3P or N2 nodes.

SemiAnalysis conducted a physical teardown of Panther Lake, Intel's latest consumer chip. It is the first commercial implementation of Backside Power Delivery Network (BSPDN) and Gate-All-Around (GAA) transistors. The teardown validates Intel's progress in advanced packaging (Foveros-S) and transistor architecture, confirming that compute cores use 18A while high-end GPU components still rely on TSMC N3E and I/O cores use TSMC N6.

Previously, market assessments relied heavily on Intel's roadmaps without physical verification. This analysis provides specific data on material choices and integration, confirming improvements in standard cell area and resistance control, while also highlighting the gap in peak density relative to Samsung SF2 and TSMC N2.

The teardown is paywalled content targeting semiconductor investors and technical experts. It offers no direct financial advice but provides key physical evidence for assessing the technical maturity of Intel's foundry business.

In his Q3 2026 update, Monevator blogger The Accumulator proposes that a passive portfolio built from scratch should split defensive assets evenly between gold, commodities, short index-linked bonds and nominal bonds, rather than relying on a single nominal-bond sleeve.

The context: Vanguard's UK Government Bond Index Fund, with roughly 13-year duration, has posted an annualised loss of 5.7% since its March 2020 peak. That drawdown dragged the Slow & Steady model portfolio to a 7.6% annualised return since launch, despite equity blocs returning 13.7% annualised over the same span.

Backtesting the alternative allocation from end-2010 yields 8.2% annualised — 0.6 percentage points higher. The mechanism is diversification of defensive sources: gold hedges currency debasement, commodities hedge supply shocks, short-duration linkers hedge real-rate rises, and nominal bonds retain liquidity buffering.

The limitation: the 0.6pp gap rests on a single backtest window dominated by one long-duration bond drawdown; a different rate or inflation regime could reverse the ranking. Four asset classes also add rebalancing complexity and tax friction. The author notes the current portfolio has only four years left to run, so the adjustment is partly driven by time horizon rather than pure return optimisation.

SoftBank founder Masayoshi Son is seeking up to $100 billion from Gulf investors to establish a new fund.

The fund aims to acquire existing companies and improve their operations using artificial intelligence and robotics to boost their value. This represents a relatively novel model for monetizing AI, differing from SoftBank's previous investment vehicles which primarily financed technology companies.

According to the Financial Times, Son has approached senior figures in the United Arab Emirates and other Middle Eastern nations. SoftBank's robotics and physical AI division, Roze, is expected to play a major role in these transformations, though exact details remain unclear.

Neither the fund's establishment nor its financing is guaranteed. It remains uncertain whether the fund will only finance acquisitions or also support the development of Roze.

In an open letter, Scott Alexander challenges Steven Pinker's view that AI intelligence is merely a set of specific problem-solving mechanisms that cannot be scaled.

Pinker argues that intelligence is inherently limited by observation and experimentation, rejecting the idea of a linear path to superintelligence. Alexander counters this by citing psychological research showing a general factor of cognition ('g') exists across species, including rats and birds.

This pattern holds for artificial systems. Research by Ruan, Maddison, and Hashimoto finds that a single general intelligence factor explains about 80% of variance in LLM performance across diverse tasks like translation and coding. Epoch's Capabilities Index further demonstrates that model performance scales predictably with parameters and data.

Alexander asserts that the default hypothesis should be that each generation of AI is more intelligent than the last, with no compelling argument for an asymptote before human-level capability. This provides an empirical basis for assessing long-term AI risks.

The main driver of recent humanoid robot progress is not hardware but specialized AI models. Tech blogger Brian Potter, writing on Construction Physics, builds up the full explanation of Vision-Language-Action (VLA) models starting from matrix multiplication.

A VLA borrows the Transformer attention mechanism from large language models, but its input is text, images, and robot sensor data, while its output is a sequence of motor commands. Figure, Unitree, Physical Intelligence, and Nvidia all use this architecture. Potter uses the open-source pi0.5 as a worked example, tracing the path from multimodal context to complex task planning.

The value for readers is a concrete framework to distinguish genuine perception-decision-execution loops from teleoperation or scripted behavior. Whether VLAs remain the dominant paradigm is uncertain, but they are the most widely used control architecture today.

The article covers public technical documentation at the principle level. It does not address proprietary performance metrics of the latest closed-source models, nor does it imply every robotics company uses the same approach.

In stochastic neural networks, categorical variables are the natural choice for representing discrete structures. However, because backpropagation cannot be performed through samples, such networks rarely use categorical latent variables. The Gumbel-Softmax estimator proposed by Jang et al. aims to address this obstacle in gradient computation.

The theoretical foundation of this method stems from the Gumbel-Max trick, which provides a simple and efficient way to draw samples from a categorical distribution with class probabilities. The authors introduce a new continuous distribution, the Gumbel-Softmax distribution, which is a continuous distribution on the simplex that can approximate categorical samples, and whose parameter gradients can be easily computed via the reparameterization trick.

A key characteristic of the Gumbel-Softmax distribution is its ability to smoothly anneal into a categorical distribution. As the softmax temperature τ approaches 0, samples from the Gumbel-Softmax distribution become one-hot encoded, and the distribution becomes identical to the categorical distribution. This interpolation capability allows models to flexibly adjust the degree of discreteness during training.

In practical applications, there is a trade-off: at low temperatures, samples are close to one-hot but have high gradient variance; at high temperatures, samples are smooth but have low gradient variance. Therefore, in practice, it is common to start at a high temperature and anneal down to a small non-zero temperature. If the temperature is treated as a learnable parameter rather than a fixed schedule, this can be interpreted as entropy regularization, allowing the distribution to adaptively adjust its "confidence".

For scenarios where discrete values must be sampled (such as action spaces in reinforcement learning), the paper proposes the Straight-Through (ST) Gumbel Estimator. It uses arg max to discretize y during the forward pass, but uses the continuous approximation to estimate gradients during the backward pass. This allows samples to maintain sparsity even at high temperatures.

Experiments show that the Gumbel-Softmax estimator outperforms existing single-sample gradient estimators in structured output prediction and unsupervised generative modeling tasks. Additionally, it can be used to efficiently train semi-supervised models, avoiding expensive marginalization operations over unobserved categorical latent variables.

"Categorical Reparameterization with Gumbel-Softmax" was written by Eric Jang, Shixiang Gu, and Ben Poole, published in 2016 (ICLR 2017). It is suitable reading for machine learning practitioners researching variational autoencoders, reinforcement learning, and any field involving optimization with discrete latent variables. It is recommended to start with Section 2 regarding the definition of the distribution and the analysis of temperature effects in Figure 1.

In his 1998 shareholder letter, Warren Buffett noted that although Berkshire's book value per share increased by 48.3%, this was primarily due to issuing stock for acquisitions. He emphasized that while such transactions instantly boost book figures, they do not result in an immediate increase in intrinsic value, because what is given up and what is received are essentially equal in nature.

This viewpoint reveals the deviation between accounting metrics and economic reality. Buffett reminded investors that what truly matters is the growth of intrinsic value per share, not the inflation of book value. When a company's stock price is significantly higher than its book value, acquiring other companies through additional stock issuance mechanically pushes up earnings per share or net assets, but this does not equate to creating new wealth for shareholders.

Furthermore, Buffett candidly discussed the constraint of scale on investment returns. He explicitly stated that as the capital base becomes excessively large, future yields will be far lower than in the past. He set a target of achieving an average annual growth rate of 15% in the future, frankly calling it "the peak that can be reached," and even acknowledged that years with negative returns could pull down the average.

For current readers, understanding how "scale effects" dilute returns is crucial. Many growth-oriented companies enjoy high compounding in their early stages, but as market capitalization expands, finding sufficiently large investment opportunities becomes difficult. Assuming a small fund achieves a 20% annualized return, its strategy may work when managing one billion dollars in assets, but become unsustainable at a ten-billion-dollar scale. This is not a degradation of ability, but a mathematical necessity.

The Berkshire Hathaway 1998 Shareholder Letter was written by Warren Buffett and published on March 1, 1999. It is suitable for investors who wish to deeply understand corporate valuation logic, beware of the trap of book profits, and recognize the limitations imposed by capital scale. It is recommended to start reading from the sections concerning "the distinction between intrinsic value and book value" and "future growth rate expectations."